Patient-reported outcomes helped predict survival in multiple myeloma using partial least squares analysis

Abstract

Objective

The prognostic value of Patient-Reported Outcomes (PRO) in predicting mortality during treatment of multiple myeloma (MM) patients was assessed using partial least square (PLS) regression, a statistical method that is well-adapted for highly correlated data.

Study Design and Setting

Four PRO measures, The European Organisation for Research and Treatment of Cancer (EORTC) QLQ-C30, the EORTC QLQ-MY24, the FACIT-Fatigue scale, and the FACT/GOG-Ntx scale, were administered during a trial designed to evaluate the efficacy and safety of bortezomib (VELCADE^® 1.3 mg/m²) in MM patients (N = 202). Clinical and PRO data were analyzed for predictive value by univariate and multivariate logistic regression methods and then by PLS regression.

Results

Fifteen baseline PRO parameters were significant in predicting mortality during treatment when univariate logistic regression was used. In contrast, only two variables were retained in the multivariate analysis, as correlated variables were excluded from the model. Using PLS regression, 14 of the 21 PRO predictors were significant in predicting mortality. Clinical and PRO data used together increased the predictive power of all models compared to clinical data alone.

Conclusion

The prognostic value of PRO was established and was more informative using PLS regression. PLS regression may therefore be a valuable method for analyzing PRO data.

Introduction

Clinical trials in cancer traditionally include quality of life (QoL) assessments “to complement the more traditional measures such as tumor response, survival, freedom from relapse and the physician's opinion concerning patient status” [1]. QoL is considered vital to measure patients' subjective tolerance to their treatment and all aspects of their well-being during and following treatment [2]. One of the largest clinical trial groups in Europe, The European Organisation for Research and Treatment of Cancer (EORTC) now include QoL measures in over 60% of all Phase III clinical trials [3]. The increase in use of QoL measures, or more generally patient reported outcomes (PRO) data, in clinical trials can also be partly attributed to the added independent prognostic value they may provide in addition to established clinical prognostic factors [4]. Given the survival benefit obtained with newer cancer therapies, the value of survival time gained, against the toxicity profile of these approaches, and their impact on the QoL of patients is therefore particularly of interest for cancer trials. Moreover, health-related quality of life (HRQL) data have recently been shown to be predictive of patient survival for multiple tumor types [5] including lung cancer and malignant melanoma [6].

When assessing the predictive value of different QoL parameters, traditional logistic regression methods are typically used, but are not ideal, as they may bias results because of 1) the exclusion of multicolinear potential predictors (highly correlated predictor variables) [7], 2) the exclusion of subjects with partial missing data (MD), and 3) the limitation of the use of a large number of variables for a small sample size. So far, in many studies, the analysis of survival and the prognostic value of QoL has been assessed using the Cox proportional hazards regression model [8], [9], [10], [11], [12]. Stepwise selection, a technique often applied with multiple and logistic regressions (to reduce a large set of predictor variables to a smaller subset deemed to be the most predictive), is also sometimes used with PRO data. However, QoL is a multidimensional concept and is usually evaluated through several variables including physical functioning, psychological well-being, social functioning, and often disease and treatment-related symptoms [13]. It is therefore important to predict survival with a statistical method that incorporates a maximum number of subjects and variables, to achieve accurate prognostic results.

Partial least square (PLS) regression, in comparison to standard linear or logistic regressions, is able to account for 1) multicolinear variables, 2) incomplete data, and 3) a larger number of variables in comparison to the number of observations. PLS regression combines features from principal component analysis (PCA) and multiple regression [14], [15], [16] and is commonly used in the field of chemometrics and genetics where the number of potential predictors can be extremely large in comparison to the number of observations [15], [16], [17]. In a PLS regression, only noninformative variables are removed and as a result, highly correlated variables remain in the model and maintain their predictive value. In addition, subjects with partial MD, which are likely critical as their data might be missing due to their health status, are also included. Excluding such patients might bias the results. Including all of these subjects and variables therefore provides a more comprehensive picture of the impact of the disease and its treatment on the patients' lives and therefore a more accurate prediction for future events.

The purpose of this study was to discuss, in more depth, the PLS methodology and its usefulness for analyzing the prognostic value of PRO data. The predictive value of baseline PRO data in predicting mortality during treatment, was analyzed through univariate and multivariate logistic regression and then by PLS regression methods. Clinical and PRO data from the SUMMIT trial were used as an illustrative example [18], [19]. Results were compared with the predictive value of baseline clinical parameters and analyzed for the added value of using PLS regression in the analysis of PRO data.